A soccer player cutting left at full speed loads their knee asymmetrically. The medial compartment compresses. The lateral compartment unloads. The stress distribution shifts in milliseconds, concentrating at points the joint was designed for straight-line walking, not lateral direction changes. The concentration points are where injuries happen.
Graph neural networks that model the joint as a mesh of interacting nodes predict these stress distributions from motion capture data (arXiv:2603.21020). Five architectures were compared: topology diffusion (information flows along mesh edges), global routing (information jumps across the mesh), and three hybrids. The hybrid combining both mechanisms — local edge-following and global long-range connections — performed best.
The superiority of the hybrid architecture reflects the physics. Stress in a loaded joint propagates locally through the cartilage matrix (topology diffusion captures this). But a force applied at the lateral condyle affects the medial compartment instantly through the rigid bone (global routing captures this). The joint has both local and global coupling simultaneously. Neither mechanism alone captures the full picture.
The clinically relevant outcome is not average stress but peak stress location. The hybrid model identified high-stress regions — the specific patches of cartilage experiencing dangerous loads — with higher fidelity than the single-mechanism models. The peaks matter more than the average because injury is a threshold phenomenon: cartilage survives moderate uniform loading but fails under concentrated peaks.
The structural insight: the knee is a system where the dangerous physics is in the spatial concentration, not the total force. An athlete who cuts harder doesn't necessarily injure themselves — one who concentrates force on a small patch does. Predicting injury requires predicting the map, not the magnitude. The topology of stress, not its total, determines the outcome.